2012 - Andrew Y. Ng - Building high-level features using large scale unsupervised learning

2012 - Andrew Y. Ng - Building high-level features using large scale unsupervised learning

ID：39350981

大小：2.92 MB

頁數(shù)：11頁

時(shí)間：2019-07-01

2012 - Andrew Y. Ng - Building high-level features using large scale unsupervised learning_第1頁
2012 - Andrew Y. Ng - Building high-level features using large scale unsupervised learning_第2頁
2012 - Andrew Y. Ng - Building high-level features using large scale unsupervised learning_第3頁
2012 - Andrew Y. Ng - Building high-level features using large scale unsupervised learning_第4頁
2012 - Andrew Y. Ng - Building high-level features using large scale unsupervised learning_第5頁
資源描述：

《2012 - Andrew Y. Ng - Building high-level features using large scale unsupervised learning》由會(huì)員上傳分享，免費(fèi)在線閱讀，更多相關(guān)內(nèi)容在學(xué)術(shù)論文-天天文庫。

1、BuildingHigh-levelFeaturesUsingLargeScaleUnsupervisedLearningQuocV.Lequocle@cs.stanford.eduMarc'AurelioRanzatoranzato@google.comRajatMongarajatmonga@google.comMatthieuDevinmdevin@google.comKaiChenkaichen@google.comGregS.Corradogcorrado@google.comJe Deanjeff@google.comAndrewY.Ngang@cs.stanford.

2、eduAbstract1.IntroductionThefocusofthisworkistobuildhigh-level,class-Weconsidertheproblemofbuildinghigh-speci cfeaturedetectorsfromunlabeledimages.Forlevel,class-speci cfeaturedetectorsfrominstance,wewouldliketounderstandifitispossibletoonlyunlabeleddata.Forexample,isitpos-buildafacedetectorfr

3、omonlyunlabeledimages.Thissibletolearnafacedetectorusingonlyunla-approachisinspiredbytheneuroscienti cconjecturebeledimages?Toanswerthis,wetraina9-thatthereexisthighlyclass-speci cneuronsinthehu-layeredlocallyconnectedsparseautoencodermanbrain,generallyandinformallyknownasgrand-withpoolingand

4、localcontrastnormalizationmotherneurons."Theextentofclass-speci cityofonalargedatasetofimages(themodelhasneuronsinthebrainisanareaofactiveinvestiga-1billionconnections,thedatasethas10mil-tion,butcurrentexperimentalevidencesuggeststhelion200x200pixelimagesdownloadedfrompossibilitythatsomeneuron

5、sinthetemporalcortextheInternet).WetrainthisnetworkusingarehighlyselectiveforobjectcategoriessuchasfacesmodelparallelismandasynchronousSGDonorhands(Desimoneetal.,1984),andperhapsevenaclusterwith1,000machines(16,000cores)speci cpeople(Quirogaetal.,2005).forthreedays.ContrarytowhatappearstoConte

6、mporarycomputervisionmethodologytypicallybeawidely-heldintuition,ourexperimentalemphasizestheroleoflabeleddatatoobtaintheseresultsrevealthatitispossibletotrainafaceclass-speci cfeaturedetectors.Forexample,tobuilddetectorwithouthavingtolabelimagesasafacedetector,oneneedsalargecollectionofimages

7、containingafaceornot.Controlexperimentslabeledascontainingfaces,oftenwithaboundingboxshowthatthisfeaturedetectorisrobustnotaroundtheface.Theneedforlargelabeledsetsposesonlytotranslationbutalsotoscalingandasigni cantchallengeforproblemsw

當(dāng)前文檔最多預(yù)覽五頁，下載文檔查看全文

此文檔下載收益歸作者所有

當(dāng)前文檔最多預(yù)覽五頁，下載文檔查看全文
溫馨提示：
1. 部分包含數(shù)學(xué)公式或PPT動(dòng)畫的文件，查看預(yù)覽時(shí)可能會(huì)顯示錯(cuò)亂或異常，文件下載后無此問題，請放心下載。
2. 本文檔由用戶上傳，版權(quán)歸屬用戶，天天文庫負(fù)責(zé)整理代發(fā)布。如果您對本文檔版權(quán)有爭議請及時(shí)聯(lián)系客服。
3. 下載前請仔細(xì)閱讀文檔內(nèi)容，確認(rèn)文檔內(nèi)容符合您的需求后進(jìn)行下載，若出現(xiàn)內(nèi)容與標(biāo)題不符可向本站投訴處理。
4. 下載文檔時(shí)可能由于網(wǎng)絡(luò)波動(dòng)等原因無法下載或下載錯(cuò)誤，付費(fèi)完成后未能成功下載的用戶請聯(lián)系客服處理。